/dev/random is not "raw entropy" it goes through the exact same mixing and (SHA-1) compression stages that /dev/urandom does. It just happens to block if there's not much entropy in the pool at the start.
I understand that, my words were not precise. My point in phrasing it that way was that there is no entropy stretching. (You only get to read (about) as much entropy as has been gathered. You don't have to worry about reading 128 bytes and having them be associated with the previously read 128 bytes, because the pools that generated them are distinct.)
Source: http://eprint.iacr.org/2012/251.pdf also, the kernel source.